tensorflow神经网络练习

import tensorflow as tf
import numpy as np

#定义隐藏层
def add_layer(inputs,in_size,out_size,activation_function=None): #默认无activation_function，则为linear_function
    Weights = tf.Variable(tf.random_normal([in_size,out_size])) #随机变量矩阵
    biases = tf.Variable(tf.zeros([1,out_size])+0.1) #因为bias推荐值不为0 
    Wx_plus_b = tf.matmul(inputs,Weights) + biases #未激活值
    if activation_function is None:
        outputs = Wx_plus_b
    else:
        outputs = activation_function(Wx_plus_b)
    return outputs

#定义输入输出及噪音
x_data = np.linspace(-1,1,300)[:,np.newaxis] #输入:300行，一个特性
noise = np.random.normal(0,0.05,x_data.shape) #正态分布normal(均值,方差,格式)
y_data = np.square(x_data)-0.5+noise#输出
#定义训练变量
xs = tf.placeholder(tf.float32,[None,1])#二维：第一维度不限，第二维度1
ys = tf.placeholder(tf.float32,[None,1])#二维：第一维度不限，第二维度1

#定义隐藏层layer1（10个）
l1 = add_layer(xs,1,10,activation_function = tf.nn.relu)#隐藏层=输入x_data,1,隐藏层数10,激活函数relu=max(0,x)
prediction = add_layer(l1,10,1,activation_function = None) #预测输出=输入为l1隐藏层，10，输出层1
#定义偏差loss=sum((实际值-预测值)^2)
loss = tf.reduce_mean(
        tf.reduce_sum(
        tf.square(ys-prediction),
        reduction_indices=[1]))
#定义训练train_step
train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss) #0.1为learning rate小于1都可以
init = tf.global_variables_initializer()
#运行输出结果
sess = tf.Session()
sess.run(init)
for i in range(1000):
    sess.run(train_step,feed_dict={xs:x_data,ys:y_data})
    if i % 50 == 0:
        print(sess.run(loss,feed_dict={xs:x_data,ys:y_data}))#打印偏差

0.313511
0.0174612
0.00860848
0.00774087
0.00724306
0.00671623
0.00615145
0.00563069
0.0050454
0.00466128
0.00435946
0.00412314
0.00394398
0.00379137
0.00365834
0.00354657
0.00345137
0.00337129
0.0033011
0.00323481